Quantitative Precipitation Forecast with a Radar Data Assimilation Method for Rainfall-induced Shallow Landslide Predictions

doi:10.21203/rs.3.rs-3802495/v1

Download PDF

Research Article

Quantitative Precipitation Forecast with a Radar Data Assimilation Method for Rainfall-induced Shallow Landslide Predictions

https://doi.org/10.21203/rs.3.rs-3802495/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Taiwan is prone to hillslope disasters in mountain areas because of its special topographical, geological, and hydrological conditions. During typhoons and rainstorms, severe shallow landslides frequently occur. To mitigate the impact of shallow landslides, not only the structural measures are necessary, but also adequate warning systems and contingency measures must be executed. Hence, precise precipitation forecasts and landslide prediction are the most important measures in practice. To account for inherent weather uncertainties precipitation forecasts based on quantitative precipitation forecast model with radar data assimilation method and ensemble mean prediction were adopted in this study instead of using a single model output. The SIMTOP model based on infinite-slope model and TOPMODEL was developed. In considering detail topographic characteristics of the subcatchments, the proposed model can estimate the variation of saturated water level during rainstorms, and then link with the slope instability analysis to clarify whether shallow landslides would occur in the subcatchment.

Two area vulnerable to landslide in Taiwan were collected to test the applicability of the model for landslide prediction. Six typhoons events were used to verify the applicability of the proposed model. Four indexes including the probability of detection (POD), false alarm ratio (FAR), threat score (TS), and accuracy (ACC) were adopted to assess the performance of two kinds of very short-term rainfall forecasts with SIMTOP model. The results indicate that quantitative precipitation forecast model with radar data assimilation with SIMTOP model was better than ensemble rainfall forecasts with SIMTOP model. It is promising to apply the proposed model for landslide early warnings to reduce the magnitude of the loss of lives and properties.

radar data assimilation method

very short-term rainfall forecast

landslide warning system

The mountainous regions of Taiwan are characterized by steep terrains and fragile geology. When typhoons hit, the intense rainfall often results in hillslope disasters, including shallow landslides and debris flow. For example, in 2015, Typhoon Soudelor caused numerous occurrences of landslide and mudflow in the mountainous areas of New Taipei City’s Xindian and Wulai Districts, cutting off the district’s only road access to the rest of the nation. In 2016, Typhoon Megi caused a shallow landslide that resulted in three deaths and the destruction of one house in Kaohsiung City’s Yanchao District. These examples reveal that the chances of disasters occurring because of heavy rainfall have increased, as have the influence levels of each disaster. Therefore, governance and planning are crucial in regions with high potential for hillslope disasters. However, because of limitations in finance and construction technology as well as land property factors, comprehensive slope protection measures cannot be implemented. To mitigate the impact of future hillslope disasters, non-structured construction measures such as the establishment of a landslide disaster warning system can be implemented. Hence, landslide early warning systems are gaining increasing attention in the literature and among government officials, decision makers, and the public (Guzzetti et al., 2020). Such landslide warning systems will reduce the time required for evacuation and resource mobilization during emergencies, thereby reducing the property loss and number of casualties caused by hillslope disasters.

In hillslope disaster prediction technology, quick and accurate quantitative precipitation forecast (QPF) is a critical input parameter. Because weather forecasts exhibit uncertain and highly nonlinear responses, QPF can be prone to errors and high uncertainty. Because of the limitations of numerical weather prediction (NWP) models and the randomness of the atmosphere, NWP cannot be error-free when only one NWP model is applied. To reduce the uncertainty of atmospheric prediction, systematic analysis and research are required for selecting the optimal combination of NWP models to develop QPF techniques. The Taiwan Typhoon and Flood Research Institute (TTFRI) and National Science and Technology Center for Disaster Reduction (NCDR) incorporated the research capacity of academia and operational agencies to conduct the Taiwan Cooperative Precipitation Ensemble Forecast Experiment (TAPEX). The TAPEX, aimed at improving the performance of QPF techniques, creates four forecasts for the following 72 hours every day (Li and Hung, 2011).

Because weather radars can acquire substantial high-resolution spatial observation information in a short time, they are applicable for producing QPFs and retrieving wind fields. Using data synthesis strategies, the acquired data can be incorporated into an atmospheric data model to obtain a more accurate initial field of the model, thereby improving its prediction performance. Therefore, the TTFRI, NCDR, Central Weather Administration, and US National Center for Atmospheric Research employed a three-dimensional (3D) variational data assimilation method (Barker et al., 2004; Gao and Stensrud, 2012) as a basis and codeveloped the radar data assimilation method. The advantage of this method is its fast calculation speed; if this speed is effectively leveraged, the method can be used to assimilate information acquired from radar observations (e.g., radar return and longitudinal wind) into a weather prediction model. This information is adopted to adjust the initial field of the atmospheric model, thereby effectively improving the QPF accuracy for 6-hour precipitation forecasts.

Studies on hillslope disaster prevention mostly rely on rainfall observation information and historical records of disaster occurrences. These studies have employed statistic regression methods to establish rainfall thresholds and rainfall intensity–delay curves; subsequently, this information is adopted to determine whether the rain would cause landsides as well as to issue landslide warnings. Caine (1980) and Hong et al. (2006) collected rain intensity–duration information in locations where mudflows had occurred and constructed a rain threshold for landslides. In 2005, the Agency of Rural Development & Soil and Water Conservation in Taiwan adopted rainfall prediction and observation information, effective cumulative rainfall depth, and rainfall triggering index to construct a warning model for rainfall-triggered mudflows. Lee et al. (2008) employed multivariate statistical models to conduct landslide potential analysis in catchment areas as well as historical typhoon events for model training and verification. Their results revealed rainfall intensity to be the key influencing factor of the model. After Typhoon Megi in 2010, the Highway Bureau in Taiwan adopted historical records of landslide disasters and rainfall to establish rainfall thresholds for landslides, thereby monitoring the key road sections and providing a reference for road closures. Because catchment areas exhibit different aggregation properties, slopes, pre-rainfall conditions, and geological and weather conditions, statistics-oriented landslide warning systems have limitations and cannot be universally applied to every catchment area. However, current advanced geographic information systems and digital elevation models (DEMs) are capable of acquiring hydrological and topographic partition of the catchment, thereby enabling the construction of precise topographic runoff model. Through further integrating the geographic information system and DEMs with slope stability analysis theories, physical landslide warning systems have been developed (Lee and Ho 2009; Park et al. 2013; Mercogliano et al. 2013; Ho and Lee 2017; Ho et al., 2022).

This study separately adopted precipitation forecasting with a radar data assimilation method and ensemble mean QPF method results as input conditions for a landslide warning system. Furthermore, it adopted the infinite slope model and TOPMODEL (SIMTOP) developed by Ho and Lee (2017) to simulate shallow landslides. The model employs the topography-based hydrological model and slope stability analysis model to simulate changes in the soil saturated water levels of a catchment and analyze the slope instability to simulate the landslide occurrence time. This study integrated the rainfall depth of 6-hour rainfall forecasts to predict the conditions of typhoons and heavy rainfall before landslides. This study aimed to quickly acquire landslide occurrence, thereby leaving sufficient response time for disaster prevention as well as mitigating the loss of lives and property.

With the aim of improving the disaster emergency response and operation abilities during typhoons and periods of heavy rain, this study separately integrated information acquired from the radar data assimilation method and ensemble QPF with the SIMTOP landslide model to develop a more comprehensive simulation calculation method for practical applications. The SIMTOP landslide analysis model employs a hydrogeomorphology model to simulate the changes in saturated water level of catchment area surface soil during periods of rainfall. After the simulation is integrated with a shallow landslide model, the SIMTOP analysis model can analyze the slope instability. Descriptions of relevant research methods are provided as follows.

2.1 Precipitation forecast with a radar data assimilation method

The advantages of using radar data include high observation durations and spatial resolutions. Past uses of radar data include making QPFs and retrieving 3D wind fields. Many studies have assimilated radar data into numeric models to develop more accurate initial fields for atmospheric models. Because a meteorological radar can quickly acquire the newest atmospheric conditions through scanning, it can provide immediate and crucial observations for very short-term (i.e., between 1 to 6 hours) QPFs. Additionally, 3D variational data assimilation employs the model forecast field of a certain period to obtain a set of fixed background error covariance matrix, which encompasses information on the relationships between control variables and their influence range. Subsequently, the observation information and background fields are adopted to establish a cost function, which is then converted into a minimal analytic model using the variational method. This analytic model is employed as the initial field model for data model forecasting, which produces more accurate model forecast results. The advantages of the 3D variational method include it requiring fewer calculation resources and being able to quickly produce assimilated results. Finally, by integrating the assimilated initial field into the atmospheric numerical model, very short-term QPF information is produced.

In 2015, the TTFRI constructed an operation standard for the wind synthesis system used for Doppler Measurement. The standard adopts 30-min rapid assimilation cycles to progressively conduct radar data assimilation. This method can be incorporated with atmospheric numerical models to instantly produce 1- to 6-hr QPF information, which serves as the input condition for landslide warning systems. Figure 1 displays the radar data assimilation strategy. In the figure, the initial background field of hour 00 is provided by the predicted field during the 18th hour. Furthermore, 3 to 4 hr of downtime are required to obtain conventional observation informationfor assimilation. Therefore, the assimilation process of the radar observation data starts from the 23rd hour. After producing forecasts from the 18th to the 23rd hour, the data model assimilates radar data acquired at 23:00, 23:30, and 00:00. After assimilating the three sets of radar data, the model is used to predict the QPF information for the following 7 hr, starting from 00 (Note: 1 hr of model calculation time must be deducted). Following this operation pattern, three sets of radar observation data are assimilated each hour to produce hourly-updated very short-term QPF information.

2.2 Ensemble QPF (TAPEX)

Because of the limitations of data-oriented weather forecast models and the randomness of the atmospheric system, data-oriented weather forecasts cannot make entirely accurate weather forecasts. By adopting statistical analysis, this study selected the optimal model combination to develop an ensemble prediction technique, thereby lowering the uncertainties of atmospheric forecasts. Because ensemble QPFs are more adaptive to weather forecast uncertainties than single-model forecasts, global and regional applications endeavor to adopt ensemble forecast systems. Furthermore, ensemble weather forecast systems can adopt the forecast results of numerous members to provide substantial and diverse forecast information, thereby improving the disadvantages of single-model forecasts. However, weather forecasts exhibit regionality and uncertainty, which obstruct the direct implementation of foreign study results. The model settings most suitable for the complex landscape of Taiwan were only obtained after conducting considerable analysis and research. Since 2010, the TTFRI has conducted the TAPEX to develop QPF technology and increase the reference information provided to disaster and emergency response agencies, thereby improving disaster prevention and rescue efficiency. The TAPEX incorporates the research capacity of academia (i.e., National Taiwan University, National Central University, National Taiwan Normal University, and the Chinese Culture University) and government (i.e., the Central Weather Bureau); integrates the computing resources from the National Center for High-Performance Computing and practical experience of the National Science and Technology Center for Disaster Reduction (NCDR) ; and adopts ensemble statistical methods and probability forecast concepts to analyze typhoon paths and rainfall distribution. One can refer to previous studies (Hsiao et al., 2012, 2013; Lee et al., 2013; Yang et al., 2015; Ho et al.,2022) for details about the TAPEX.

Four high-resolution weather forecasts for Taiwan in the following 78 hours (6 hr of atmospheric model computing time must be deducted), indicating that four precipitation forecast results for the following 72 hours can be produced daily. This study applied 20 ensembles of QPF data from the TAPEX data weather forecast model to compute 6-hr forecasts using the ensemble mean precipitation forecast method. Additionally, through integrating real-time precipitation observation information from physical rainfall stations, the model can conduct real-time rolling updates on the precipitation forecast information for the following 6 hours. This information can be adopted in the forecast analysis of shallow landslide warning systems.

2.3 SIMTOP model

This study employed the SIMTOP model proposed by Ho and Lee (2017) to analyze the stability of catchment areas. The model integrates simulated soil saturated water level and infinite slope stability analysis to calculate the factor of safety of a catchment area and determine whether shallow landslides may occur there. Conventional infinite slope stability analysis adopts grid-based computing methods and employs the factor of safety (FS) to determine stable and unstable grids. SIMTOP model proposed adopting a catchment area’s mean factor of safety $\left(\overline{FS}\right)$ to evaluate whether shallow landslides may occur there. The analyzing formula is expressed as follows:

$$\overline{FS}\left(t\right)=\frac{1}{N}{\sum }_{j=1}^{N}\left(\frac{{C}_{j}+\left({\rho }_{s}{D}_{j}-{\rho }_{w}{h}_{\text{w,}j}\left(t\right)\right)g{{cos}}^{2}{\beta }_{j}{tan}{\varphi }_{j}}{{\rho }_{s}g{sin}{\beta }_{j}{cos}{\beta }_{j}}\right)$$

where N represent the number of grid of the catchment area; ${C}_{j}$ represents the effective cohesive force [N/m²] at $j$ grid, ${\rho }_{s}$ is the soil bulk density [kg/m³], ${\rho }_{w}$ is the water density [kg/m³], ${\beta }_{j}$ indicates the inclination angle of the hillslope at $j$ grid, ${\varphi }_{j}$ represents the soil effective friction angle $j$ grid, $g$ is the gravitational acceleration (= 9.81 m/s²), ${D}_{j}$ is the soil thickness [m], and ${h}_{\text{w,}j}\left(t\right)$ indicates the saturated water level at the time t and $j$ grid [m]. Eq. (1) indicates that the $\overline{FS}\left(t\right)$ varies temporally along with changes in the saturated groundwater level. Additionally, $\overline{FS}\left(t\right)$<1 represents that at time t, the comprehensive slope driving force surpasses the resistance force; therefore, the catchment area exhibits unstable slopes. By contrast, $\overline{FS}\left(t\right)\ge 1$ indicates that at t, the comprehensive driving force large than the resistance force, resulting in stable slopes within the catchment area.

This study selected the catchment areas of (a) Chung-Chih village catchment (C1) in New Taipei City’s Wulai District, and (b) Dong-Yen village catchment (C2) in Kaohsiung City’s Yan-Chao District as model testing areas. Furthermore, it adopted the shallow landslide events that occurred between January 1, 2010 and October 15, 2017 for follow-up analysis. The model’s application and analysis mainly comprised the following three parts: (1) through adopting six typhoon events as testing examples, this study applied the QPF with radar data assimilation method and ensemble mean precipitation forecast method to produce 6-hr precipitation forecasts and analysis. (2) This study employed the observed rainfall depth as the input condition for the SIMTOP landslide model and conducted a long-term systematic evaluation to test the accuracy of the SIMTOP model. (3) It integrated very short-term precipitation forecasts into the SIMTOP landslide model to analyze the model’s warning performance for the next 6 hours.

3.1 Description of study areas

(a) Chung-Chih village catchment (C1)

In 2015, Typhoon Soudelor devastated Wu-Lai in northern Taiwan and caused severe shallow landslides and mudflows occurred on No. 9A Highway at 10.2 K. This study referenced the disaster record information of the Highway Bureau and Agency of Rural Development and Soil Conservation in Taiwan. Subsequently, Chung-Chih Village catchment (C1) is selected to the study area. Figure 2 displays the geographic location of said slope area in the Nan-Shih river basin. This study selected the Shang-Gui-Shan Bridge flow gauge station in the basin to verify the parameters of the SIMTOP model.

Chung-Chih village catchment (C1) is on the No. 9A Highway at 10.2 K (Fig. 2) and adopted 5-m resolution DEM. The DEM analysis results revealed that the catchment area was 0.59 km² and the average slope angle was 35.9°. According to the traffic block information of the Directorate General of Highways displayed, three severe shallow landslide disaster events occurred on the No. 9A Highway at 10.2 K between 2010 and 2017, which were caused by Typhoons Soudelor (2015), Dujuan (2015), and Megi (2016). In the follow-up analysis, this study adopted the traffic congestion times on No. 9A Highway as the landslide occurrence times.

(b) Dong-Yen village catchment (C2)

In 2016, strong wind and heavy rainfall caused by Typhoon Megi induced numerous disasters in Taiwan. The most notable hillslope disaster occurred on September 28 when a shallow landslide occurred in Dong-Yen village catchment (C2) in in Kaohsiung City. Approximately 8,700 m³ of soil slid for approximately 20 m, burying one house and three persons were killed.

Dong-Yen village (C2) is located close to the Yen-Feng bridge watershed (Fig. 3). The Yen-Feng bridge watershed is 17 km². The detail rainfall and flow discharge data of the Yeng-Feng Bridge watershed were also used to verify the hydrological parameters of SIMTOP model. C2 area is 0.58 km² and average slope is 0.574. According to the 1/250,000 scale geological map from the Geological Survey and Mining Management Agency in Taiwan, the area is composed of sandstone, mud rock, and shale.

3.2 Analysis and discussion of rainfall forecasts

As shown in Table 1, this study first evaluated the accuracy of the rainfall forecast results and adopted six typhoon-induced heavy rainfall events for analysis. It separately applied the QPF with radar data assimilation method and ensemble mean precipitation forecast method with real-time rainfall observation information from the physical rainfall stations to produce rolling updates of 1- to 6-hr forecast information. Furthermore, this study employed the coefficient of efficiency (CE_R) and coefficient of correlation (CC_R) to verify the feasibility of adopting precipitation forecasts to predict the cumulative rainfall. The calculation of CE_R is displayed as follows:

Table 1

Storm events and landslide occurrence records analyzed in this study
Typhoon Event	Warning period	Duration (hr)	Landslide occurrence
Megi	2016-09-25 23:30 2016-09-28 17:30	66	C1、C2
Meranti	2016-09-12 23:30 2016-09-15 11:30	36	-
Nepartak	2016-07-06 14:30 2016-07-09 14:30	72	-
Dujuan	2015-09-27 08:30 2015-09-29 17:30	57	C1
Soudelor	2015-08-06 11:30 2015-08-09 08:30	71	C1
Linfa	2015-07-06 08:30 2015-07-09 05:30	69	-

Table 2

Geomorphologic factors of the study watersheds.
Watershed	i	N_i	${\overline{A}}_{i}$ (km²)	${{\overline{L}}_{c}}_{i}$ (km)	${\overline{S}}_{{o}_{i}}$ (m/m)	${\overline{S}}_{{c}_{i}}$ (m/m)
Shang-Gui-Shan Bridge (C1)	1	680	0.28	0.45	0.6656	0.3456
	2	145	1.29	1.03	0.7154	0.2290
	3	33	6.34	2.72	0.7170	0.0947
	4	9	24.06	4.31	0.6869	0.0415
	5	3	72.48	7.52	0.7268	0.0169
	6	1	334.83	23.82	0.6847	0.0121
Yan-Feng Bridge (C2)	1	8	1.27	1.46	0.3218	0.0173
	2	2	5.17	2.34	0.2406	0.0062
	3	1	16.68	7.84	0.1483	0.0019
Notes: N_i is the number of i^th-order streams; ${\overline{A}}_{i}$ is the mean ith-order subwatershed area; ${{\overline{L}}_{c}}_{i }$is the mean i^th-order channel length; ${\overline{S}}_{{o}_{i}}$ is the mean i^th -order hillslope slope; ${\overline{S}}_{{c}_{i}}$ is the mean i^th -order channel slope.

$${\text{CE}}_{\text{R}}\text{=1}-\frac{\sum _{\text{t}\text{=}\text{1}}^{\text{n}}{\left({\text{RA}}_{\text{t}}-{\widehat{\text{RA}}}_{\text{t}}\right)}^{\text{2}}}{\sum _{\text{t}\text{=}\text{1}}^{\text{n}}{\left({\text{RA}}_{\text{t}}-{\stackrel{\text{-}}{\text{RA}}}_{\text{t}}\right)}^{\text{2}}}$$

where RA_t, ${\widehat{\text{RA}}}_{\text{t}}$, and $\overline{RA}$ represent the historical records of the cumulative rainfall at time t, the predicted cumulative rainfall at time t, and the average of recorded cumulative rainfall, respectively. A CE_R value closer to 1 demonstrates greater forecast accuracy. Additionally, the calculation of CC_R is displayed as follows:

$${\text{CC}}_{\text{R}}\text{=}\frac{\sum _{\text{t=1}}^{\text{n}}\left({\text{RA}}_{\text{t}}-\stackrel{\text{-}}{\text{RA}}\right)\left({\widehat{\text{RA}}}_{\text{t}}-\stackrel{\text{-}}{\widehat{\text{R}}\text{A}}\right)}{\sqrt{\sum _{\text{t=1}}^{\text{n}}{\left({\text{RA}}_{\text{t}}-\stackrel{\text{-}}{\text{RA}}\right)}^{\text{2}}\text{∙}\sum _{\text{t=1}}^{\text{n}}{\left({\widehat{\text{RA}}}_{\text{t}}-\stackrel{\text{-}}{\widehat{\text{R}}\text{A}}\right)}^{\text{2}}}}$$

where $\stackrel{\text{-}}{\widehat{\text{R}}\text{A}}$ represents the mean of the predicted cumulative rainfall. A CC_R value closer to 1 represents greater forecast accuracy.

This study selected two rainfall stations that are close to the research areas, namely those at Da-Tung-Shan rainfall station in C1, and Heng-Shan Elementary School stations in C2, for coefficient of efficiency (Eq. (2)) and correlation (Eq. (3)) analyses. Figures 4 and 5 display the analyses’ results. The correlation of efficiency (CE_R) and coefficient of correlation (CC_R) results indicated that the 1-hr to 3-hr forecasts of the two forecast methods were adequate. However, the two methods started to yield considerably different results in 4-hr to 6-hr. A possible cause of this disparity was the growing randomness of rainfall, which made rainfall development or recession increasingly difficult to predict as forecast time lengthened. By adopting the radar data assimilation method to constantly update the initial forecast background field, the progressive 6-hr precipitation forecast accuracy of the very short-term precipitation forecast was improved. Furthermore, the coefficient of correlation results of the QPF with radar data assimilation method signified a high correlation between actual and predicted cumulative rainfalls. Therefore, the rainfall characteristics forecast of this method was mostly adequate. Furthermore, the 6-hr precipitation forecasts of the six typhoons events revealed that the accuracy of this method was greater than that of the ensemble mean precipitation forecast method.

This study adopted Typhoon Soudelor from 2015 as an example (Fig. 6) and selected the duration between warning period as the analysis period. In Fig. 7, the horizontal and vertical axes represent the observed and predicted cumulative rainfall depth, respectively. When the predicted cumulative rainfall matches the observed cumulative rainfall, the curve has a gradient of 45°. This study compared the observed cumulative rainfall of the rainfall station with the 1-hr to 6-hr forecasts of the cumulative rainfall values produced by the two forecast methods. The results indicate that the cumulative rainfall prediction of the QPF with radar data assimilation method is adequate. In comparison, the cumulative depth prediction of the ensemble mean precipitation forecast method often underestimated the accumulated rainfall. Subsequently, the two precipitation methods are separately integrated with the SIMTOP model to compute the 1-hr to 6-hr forecasts of the catchment area’s mean FS.

3.2 SIMTOP model’s parameter calibration

Because the SIMTOP model required numerous geomorphological factors of the catchment area, this study adopted high-resolution 5 × 5 m DEM for computation. The method of Lee (1998) was employed to construct a DEM for acquiring the required geomorphological factors for the hydrographical model, including the watershed area, river lengths, and average slope angles of the catchment area overland and channel flows. Data from the Shang-Gui-Shan Bridge station in C1 and Yan-Feng stations were adopted to evaluate and verify the model variables. Figures 2 & 3 display the geographic locations of the rainfall stations. Table 1 shows the geomorphologic factors of the watersheds.

To perform runoff simulation of typhoon and flooding events in the SIMTOP model, this study conducted hourly rainfall-runoff simulation and analysis of the Shang-Gui-Shan Bridge and Yan-Feng Bridge watershed during different typhoon events. The analysis results revealed that the recession parameter (m) of the sample catchment area was 0.06 m; the hydraulic conductivity (K₀) of the saturated soil was 5×10^–3 m/s²; and the maximum storage in the root zone (S_RZmax) was 0.02 m. The variables required for infinite slope stability analysis are accessible through on-site soil ampling and digital elevation model. This study set the soil friction angle as 30°, the effective cohesion as 2000 kN/m², and the soil bulk density as 19.62 kN/m³. Many studies (Montgomery et al., 1994; Casadei et al., 2003; Ho and Lee, 2017; Ho et al., 2022) have indicated that these soil parameters are within a reasonable range. Please refer Ho et al. (2022) for more details on rainfall-runoff simulation and SIMTOP model’s landslide prediction. This study employed this set of parameters in the SIMTOP model to conduct long-term systematic evaluations of the selected catchment areas.

3.3 Integrating precipitation forecasts with SIMTOP landslide model analysis

This study analyzed the Chung-Chih village catchment (C1) and Dong-Yen village catchment (C2) during the interval between the announcement and dismissal of the offshore warning for Typhoon events. The QPF with radar data assimilation method and ensemble mean precipitation forecast method were separately integrated with the SIMTOP landslide model to compute factor of safety variation for the next 6-hr at each hour. Figure 7 indicates that when making forecasts on typhoon events, the peak rainfall occurrence time prediction of the QPF with radar data assimilation method was largely accurate. However, the method slightly overestimated the total rainfall. Alternatively, ensemble mean precipitation forecast method produced comprehensive precipitation type forecasts that were occasionally delayed compared with historical records and underestimated the total rainfall. By integrating the QPF with radar data assimilation method into the SIMTOP model, the proposed model became capable of generating effective landslide warnings as well as demonstrated greater landslide prediction performance than the SIMTOP model integrated with the ensemble mean precipitation forecast method.

This study also adopted Typhoon Soudelor from 2015 as an example and selected the duration between warning period as the analysis period to show the integrating precipitation forecasts with SIMTOP landslide model analysis. In Fig. 8, the blue line represented the and mean FS of this catchment using the historical hourly rainfall, and, black line with six red dots (6-hr prediction) represented mean FS of this catchment using the precipitation forecasts. Furthermore, the mean FS values changed in accordance with rainfall intensity and cumulated rainfall depth. This is because increased rainfall results in a greater amount of water infiltrating the soil. When the accumulated rainfall exceeds the infiltration capacity, the saturated water level gradually rises, and then the possibilities of shallow landslides will increase. The solid and hollow red arrows represent the landslide events that did and did not occur within the catchment area, respectively.

To compare the model prediction results with the simulated results, an error matrix was adopted to determine the accuracy of each model. As shown in Table 3, the error matrix produces four types of result conditions, namely hits, false alarms, misses, and no events, which refer to the model accurately predicting a simulated landslide occurrence, predicting a landslide occurrence despite no simulated landslide occurring, failing to predict a simulated landslide occurrence, and accurately predicting the simulated non-occurrence of a landslide, respectively. Finally, the frequency or time-sequence FS of each result condition was compared to establish the following variables to evaluate the suitability of the integrating precipitation forecasts with SIMTOP landslide model: probability of detection (POD), false alarm ratio (FAR), threat score (TS), and accuracy (ACC). The computing equations are listed as follows (Wilks 2011; Schaefer 1990):

$\text{POD=}\frac{\text{Hits}}{\text{Hits+Misses}}$	(4)
$\text{FAR}=\frac{\text{False alarms}}{\text{Hits}+\text{False alarms}}$	(5)
$\text{TS}=\frac{\text{Hits}}{\text{Hits}+\text{False alarms}+\text{Misses}}$	(6)
$\text{ACC}=\frac{\text{Hits}+\text{No events}}{\text{Hits}+\text{False alarms}+\text{Misses}+\text{No events}}$	(7)

Table 3

Error matrix
simulation prediction	$\overline{FS}$<1	$\overline{FS}\ge 1$
$\overline{FS}\ge 1$	Misses	No events
$\overline{FS}$<1	Hits	False alarms

The POD, as presented in Eq. (4), ranges between 0 and 1 and represents the proportion of successful predictions of shallow landslide events. A POD value closer to 1 indicates higher model prediction performance. However, because the POD is derived by considering only hits while ignoring false alarms, a smaller sample size may generate an overstated POD value. Therefore, the FAR is typically included in performance assessments. The FAR, as presented in Eq. (5), represents the proportion of false alarms for an event under evaluation. An FAR value closer to 0 indicates higher model prediction performance. The TS, as presented in Eq. (6), ranges between 0 and 1 and is evaluated by considering both false alarms and misses; this index can be considered as representing the model prediction accuracy. The TS value closer to 1 indicates higher model prediction performance; the TS value is equal to 0 when the model exhibits no predictive capability. Finally, the ACC, as presented in Eq. (7), ranges between 0 and 1 and represents the all of successful and unsuccessful predictions of shallow landslide events. A ACC value closer to 1 indicates higher model prediction accuracy.

Figure 9 displays the error matrix analysis (Table 3) of the simulated mean FS and prediction results. The results indicated that the 1- to 4-hr mean FS predictions of the model were highly accurate; however, the 5- to 6-hr predictions of the mean FS barely achieved the required accuracy. This is because longer prediction times resulted in greater randomness of rainfall, which made rainfall development or recession difficult to predict. Therefore, acquiring accurate forecasts for longer forecast times is difficult.

To thoroughly understand the landslide prediction performance of this forecast model in the C1 and C2 during various typhoon events, this study further adopted six typhoons (Megi, Meranti, and Nepartak in 2016 and Dujuan, Soudelor, and Linfa in 2015) for testing as shown in Table 1. Table 4 displays the analysis results. Regarding the 1- to 4-hr landslide prediction performance of the radar data assimilation method, its POD, FAR, TS, and ACC values were 0.813, 0.480, 0.131, and 0.979, respectively, which were adequate prediction results. However, longer forecast times result in greater variations in precipitation forecasts. Furthermore, POD, FAR, TS, and ACC values of the 5- to 6-hr forecasts were reduced to 0.625, 0.545, 0.357, and 0.974, respectively. By contrast, in terms of the 1- to 4-hr landslide forecasts using the ensemble mean precipitation forecast method, the POD, FAR, TS, and ACC values were 0.412, 0.533, 0.280, and 0.974, respectively. The analysis results revealed that regardless of which QPF method was integrated into the landslide prediction evaluation model, prediction performance deteriorated as forecast time increased. Nevertheless, the combination of QPF with radar data assimilation method and with the SIMTOP model landslide warning system outperformed the combination of ensemble QPF and the SIMTOP model landslide warning system. Therefore, this warning system is suitable for generating 1- to 6-hr predictions of slope stability, and thus, can win more response time and reduce the loss of lives and property.

Table 4

Results of landslide prediction using two kinds of precipitation forecasts
Method	Ahead time	Misses	No events	Hits	False alarms	POD	FAR	TS	ACC
QPF with radar data assimulation	1 hr	0	724	16	1	1.000	0.059	0.941	0.999
	2 hr	0	712	16	1	1.000	0.059	0.941	0.999
	3 hr	1	691	15	10	0.938	0.400	0.577	0.985
	4 hr	3	677	13	12	0.813	0.480	0.464	0.979
	5 hr	6	666	10	11	0.625	0.524	0.370	0.975
	6 hr	6	654	10	12	0.625	0.545	0.357	0.974
Ensemble mean	1 hr	1	723	16	1	0.941	0.059	0.889	0.997
	2 hr	1	711	16	1	0.941	0.059	0.889	0.997
	3 hr	7	696	10	4	0.588	0.286	0.476	0.985
	4 hr	10	680	7	8	0.412	0.533	0.280	0.974
	5 hr	13	669	4	7	0.235	0.636	0.167	0.971
	6 hr	14	656	3	9	0.176	0.750	0.115	0.966

This study separately applied the QPF with radar data assimilation method and the ensemble mean precipitation forecast method to make 1- to 6-hr precipitation forecasts in catchment areas. By integrating these forecast methods with the SIMTOP shallow landslide warning system for model analysis, a well-structured shallow landslide immediate warning system was constructed. Additionally, six typhoon events were adopted for testing to thoroughly determine the landslide prediction performance of this model on different typhoon events. The 1- to 4-hr landslide prediction performance of the QPF with radar data assimilation method achieved POD, FAR, TS, and ACC values of 0.813, 0.480, 0.464, and 0.979, respectively, indicating adequate prediction results. However, in terms of the 5- to 6-hr landslide prediction results, the POD, FAR, TS, and ACC values dropped to 0.625, 0.545, 0.357, and 0.974, respectively. By contrast, regarding the 1- to 4-hr landslide prediction performance of the ensemble mean precipitation forecast method, the POD, FAR, TS, and ACC values were 0.412, 0.533, 0.280, and 0.974, respectively. These results indicated that the landslide warning evaluation performance of both integrated models exhibited reduced prediction performance as the forecast time increased. Further analysis revealed that after integrating the QPF with radar data assimilation method into the SIMTOP model landslide warning system, the system exhibited slightly stronger landslide prediction performance compared with ensemble QPF. Therefore, the QPF with radar data assimilation method-integrated SIMTOP landslide prediction model developed in this study can effectively predict the timing and location of potential landslides during typhoons, thereby enhancing immediate emergency response and operation abilities during typhoon or flooding periods. This study aimed to assist relevant disaster prevention management agencies in taking suitable precautions and establishing response measures to reduce losses caused by intense typhoon rainfall.

Ethics approval and consent to participate

The authors would like to confirm that they have no competing of interest regarding this paper. All authors worked closely on this project starting from the conception of the project through modelling, analysis, interpretation of the results and write up of the manuscript.

Consent for publication

Not applicable.

Availability of data and materials

Vs data will be shared through the Agency of Rural Development and Soil and Water Conservation, Taiwan, after this paper is published.

Funding

This research was supported by the Agency of Rural Development and Soil and Water Conservation, Taiwan, under grant SWCB-105-159. The authors also acknowledge the financial support provided by the National Science and Technology Council in Taiwan.

Competing interests

The authors declare that they have no competing interests

Authors’ Contributions

Jui-Yi Ho,Chung-Yi Lin, Che-Hsin Liu, Wei-Bo Chen, and Chih-Hsin Chang were validated proposed model. Kwan Tun Lee was designed and performed the model. All authors contributed about equally to compose this paper, discussed the results, and commented on the manuscript at all stages.

Acknowledgments

The authors would like to thank the Taiwan Typhoon and Flood Research Institute and National Science and Technology Center for Disaster Reduction, Taiwan, for providing the rainfall forecasts and the Central Weather Administration, Taiwan, for providing rainfall observation data.

Barker DM, Huang W, Guo YR, Bourgeois AJ, Xiao QN (2004) A three-dimensional variational data assimilation system for MM5: Implementation and initial results. Mon Weather Rev 132(4):897–914
Beven KJ, Kirkby MJ (1979) A physically based, variable contributing area model of basin hydrology. Hydrol Sci J 24(1):43–69
Caine N (1980) The rainfall intensity: duration control of shallow landslides and debris flows. Geogr Ann Ser Phys Geogr : 23–27
Casadei M, Dietrich WE, Miller NL (2003) Testing a model for predicting the time and location of shallow landslide initiation in soil-mantled landscapes. Earth Surf Processed Land 28:925–950
Gao J, Stensrud DJ (2012) Assimilation of reflectivity data in a convective-scale, cycled 3DVAR framework with hydrometeor classification. J Atmos Sci 69(3):1054–1065
Guzzetti F, Gariano SL, Peruccacci S, Brunetti MT, Marchesini I, Rossi M, Melillo M (2020) Geographical landslide early warning systems. Earth–Science Reviews 200:102973
Ho J-Y, Lee KT (2017) Performance evaluation of a physically based model for shallow landslide prediction. Landslides 14(3):961–980
Hong Y, Adler R, Huffman G (2006) Evaluation of the potential of NASA multi-satellite precipitation analysis in global landslide hazard assessment. Geophys Res Lett 33(22):L22402
Ho J-Y, Lee KT (2022) Using ensemble quantitative precipitation forecast for rainfall-induced shallow landslide predictions. Geoscience Letter 9(22):1–13
Lee CT, Huang CC, Lee JF, Pan KL, Lin ML, Dong JJ (2008) Statistical approach to storm event-induced landslide susceptibility. Nat Hazard Earth Syst Sci 8:941–960
Lee KT (1998) Generating design hydrographs by DEM assisted geomorphic runoff simulation: a case study. J Am Water Resour Assoc 34(2):375–384
Lee KT, Ho J-Y (2009) Prediction of landslide occurrence based on slope-instability analysis and hydrological model simulation. J Hydrol 375(3):489–497
Li J-S, Hong J-S (2011) The study of regional ensemble forecast: Physical perturbations. Atmospheric Sci 39(2):95–115
Mercogliano P, Casagli N, Catani F, Damiano E, Olivares L, Picarelli L, Tofani V (2013) Short term weather forecasting for shallow landslide prediction. Landslide Science and Practice. Springer, Berlin, Heidelberg
Montgomery DR, Dietrich WE (1994) A physically based model for the topographic control on shallow landsliding. Water Resour Res 30(4):1153–1171
Park HJ, Lee JH, Woo I (2013) Assessment of rainfall-induced shallow landslide susceptibility using a GIS-based probabilistic approach. Eng Geol 161:1–15
Schaefer JT (1990) The critical success index as an indicator of warning skill. Weather Forecast 5(4):570–575
Wilks DS (2011) Statistical methods in the atmospheric sciences. Academic press

Download PDF

Reviewers agreed at journal
09 Jan, 2024
Reviewers invited by journal
09 Jan, 2024
Editor invited by journal
08 Jan, 2024
Editor assigned by journal
27 Dec, 2023
First submitted to journal
26 Dec, 2023

You are reading this latest preprint version

Quantitative Precipitation Forecast with a Radar Data Assimilation Method for Rainfall-induced Shallow Landslide Predictions

Status:

Version 1

Abstract

Figures

1. Introduction

2. Methodology

2.1 Precipitation forecast with a radar data assimilation method

2.2 Ensemble QPF (TAPEX)

2.3 SIMTOP model

3. Model Application and Analysis

3.1 Description of study areas

3.2 Analysis and discussion of rainfall forecasts

3.2 SIMTOP model’s parameter calibration

3.3 Integrating precipitation forecasts with SIMTOP landslide model analysis

4. Conclusion

Declarations

References

Status:

Version 1

\(\text{POD=}\frac{\text{Hits}}{\text{Hits+Misses}}\)	(4)
\(\text{FAR}=\frac{\text{False alarms}}{\text{Hits}+\text{False alarms}}\)	(5)
\(\text{TS}=\frac{\text{Hits}}{\text{Hits}+\text{False alarms}+\text{Misses}}\)	(6)
\(\text{ACC}=\frac{\text{Hits}+\text{No events}}{\text{Hits}+\text{False alarms}+\text{Misses}+\text{No events}}\)	(7)

simulation prediction	\(\overline{FS}\)<1	\(\overline{FS}\ge 1\)
\(\overline{FS}\ge 1\)	Misses	No events
\(\overline{FS}\)<1	Hits	False alarms